Librarian

What it is

A Dockerized agent for an Obsidian vault. It watches a drop folder for emails, chat exports and web clippings, files them as linked pages, vectorizes everything, and answers questions with citations. It sends one sentence to a master agent through Agent Update only when something matters.

The problem

Notes and documents pile up unfiled, and nobody maintains the vault or can query it reliably.

How it works

1. A raw/ drop folder is watched; the agent reads each item, works out what it connects to and writes it as a proper page
2. Retrieval uses local model2vec embeddings plus BM25 fused with reciprocal rank fusion, so there is no GPU or per-token embedding cost
3. Reasoning runs through codex exec on an existing ChatGPT plan, with separate larger and smaller models for filing and triage
4. A Python Sandbox class and the container both enforce the vault boundary, and a heartbeat leash stops the agent if the terminal session dies
5. Terminal dashboard with a baseline mode that marks existing history as already filed

Stack

Disciplines

For your business

A resident knowledge-base agent that files incoming documents, keeps them searchable and answers questions with citations.